OPEN LINK ↗
// 1h agoRESEARCH PAPER
Turn-BoN scales LLM agents at inference
This paper evaluates turn-level best-of-N sampling, where a generator proposes multiple actions and a selector chooses one before execution. On Terminal-Bench v2.1, the approach improves accuracy and consistency while reducing agent turns, but requires substantially more inference compute.
// ANALYSIS
Turn-BoN offers a compelling way to improve agent reliability without retraining models, though its 3–6× token cost makes deployment challenging.
- –Improves task accuracy by roughly 5–10% across long-horizon terminal tasks
- –Cuts average turns to completion by approximately 40–50%
- –Encourages better debugging, testing, instruction following, and fewer loops
- –Uses off-the-shelf open-weight models as selectors instead of a dedicated critic
- –Compute overhead remains the central obstacle, especially at larger sampling widths
// TAGS
turn-bonagentinferenceevaluationbenchmarkreasoning
DISCOVERED
1h ago
2026-11-05
PUBLISHED
2h ago
2026-11-05
RELEVANCE
9/ 10
AUTHOR
arnaudmercier